[https://nvbugs/6163033][fix] Guard q_a_proj.weight dict access behind nvfp4_fused_a; update test to `chec

tensorrt_llm/_torch/models/modeling_deepseekv3.py

@@ -440,8 +440,8 @@ def detect_shared_mtp_weights() -> bool:
                     # Non-lite models (V3, R1, V3.2) fuse q_a_proj into
                     # kv_a_proj_with_mqa, so both must be NVFP4 for the fused
                     # path. Lite models (V3-Lite) have no q_a_proj.
-                    if not is_lite:
-                        nvfp4_fused_a &= weights[
+                    if not is_lite and nvfp4_fused_a:
+                        nvfp4_fused_a = weights[
                             f"{'.'.join(names[:-1])}.q_a_proj.weight"].dtype == fp4_utils.float4_e2m1x2
                     if nvfp4_fused_a:
                         ########### input_scale

tensorrt_llm/_torch/models/modeling_mistral.py

@@ -310,13 +310,19 @@ def get_num_tokens_per_image(self, image_size):
             Image.new("RGB", (w, h)))
         return ncols * nrows + nrows
 
-    def __call__(self, text, images, **kwargs):
-        mm_items = []
-        if images:
-            mm_items = [{
-                "type": "image",
-                "base64": encode_base64_image(image)
-            } for image in images]
+    def __call__(self, text, images=None, **kwargs):
+        if not images:
+            # Plain-text inputs (e.g. text-only evaluation like MMLU/GSM8K): tokenize
+            # directly without wrapping in a multi-modal chat conversation, which would
+            # otherwise inject chat-template tokens and corrupt continuation prompts.
+            encoded = self.tokenizer.transformers_tokenizer(text,
+                                                            return_tensors='pt')
+            return {"input_ids": encoded["input_ids"]}
+
+        mm_items = [{
+            "type": "image",
+            "base64": encode_base64_image(image)
+        } for image in images]
 
         conversation = [{
             "role": "user",

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -6784,7 +6784,7 @@ def test_fp8(self, tp_size, pp_size, ep_size, attention_dp, cuda_graph,
                 eagle3_model_arch="mistral_large3")
         with LLM(
                 f"{llm_models_root()}/Mistral-Large-3-675B/Mistral-Large-3-675B-Instruct-2512/",
-                checkpoint_format="mistral",
+                checkpoint_format="mistral_large_3",
                 tensor_parallel_size=tp_size,
                 pipeline_parallel_size=pp_size,
                 moe_expert_parallel_size=ep_size,

tests/integration/test_lists/waives.txt

@@ -161,7 +161,6 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_tp4[torch_compile=True] SKIP (https://nvbugs/5821415)
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_nvfp4_tp4[torch_compile=True] SKIP (https://nvbugs/5821415)
 accuracy/test_llm_api_pytorch.py::TestMiniMaxM2::test_4gpus[attention_dp=False-cuda_graph=True-overlap_scheduler=True-tp_size=4-ep_size=4] SKIP (https://nvbugs/6159132)
-accuracy/test_llm_api_pytorch.py::TestMistralLarge3_675B::test_fp8[latency_moe_deepgemm] SKIP (https://nvbugs/6163033)
 accuracy/test_llm_api_pytorch.py::TestMistralLarge3_675B::test_nvfp4_4gpus[latency_moe_trtllm] SKIP (https://nvbugs/6162121)
 accuracy/test_llm_api_pytorch.py::TestMistralLarge3_675B::test_nvfp4_4gpus[latency_moe_trtllm_eagle] SKIP (https://nvbugs/6157892)
 accuracy/test_llm_api_pytorch.py::TestPhi4MiniInstruct::test_auto_dtype SKIP (https://nvbugs/6076767)